Optimasi Analisis Sentimen Ulasan Platform Pendidikan Daring Menggunakan Arsitektur ALBERT dan Teknik Augmentasi Kontekstual

Abstract

Online learning through global platforms like Coursera generates a massive volume of user reviews, which serve as vital information for educational quality improvement. However, these reviews often exhibit imbalanced label distributions, where positive sentiments significantly dominate negative and neutral ones, hindering traditional classification models. Advanced language models such as ALBERT (A Lite BERT) offer parameter efficiency through cross-layer parameter sharing while maintaining high performance in complex text understanding. This study aims to evaluate the ALBERT model's performance in classifying Coursera user reviews and addressing data imbalance using Contextual Word Embedding augmentation. The methodology involves collecting 10,000 reviews followed by preprocessing steps including case folding, punctuation removal, and tokenization. The augmentation technique utilizes language models to replace words based on context to balance minority classes. The results show that ALBERT provides highly consistent performance, achieving an F1-score of 0.9710 with the contextual augmentation scenario. The model proves effective in capturing linguistic variations and remains computationally efficient. In conclusion, the ALBERT model is highly effective for sentiment analysis on the Coursera dataset, where contextual augmentation significantly enhances the model's ability to recognize minority classes that were previously difficult to identify.